Automatic Modulation Classification based on Bayesian Learning

This study applies Bayesian learning techniques, specifically Variational Inference (VI) and Monte Carlo Dropout (MC Dropout) to Automatic Modulation Classification (AMC). Both methods are built upon a Convolutional Neural Network (CNN) framework and are capable of detecting certain out-of-domain (OOD) modulations through threshold- based decision making. The results indicate that the confidence scores assigned to OOD modulations are influenced by their proximity to the trained in-domain classes. Higher confidence scores are typically observed for modulation types that closely resemble known modulation classes. This approach enables effective classification by training only on essential modulation types, thereby reducing the need for exhaustive class coverage and supporting the constraints of edge computing environments.